p.s. in this demo the AI replies using this canned response!

Philippe Lehoux: What would be a game changer in the medium term would be to provide the AI with ALL the previous messages in the current thread, or at least as many as possible within the max context window, leaving some configurable minimal size for the response. This will especially make things much more relevant when we can use a 32k tokens context window. But OpenAI will have to drop its pricing as otherwise it'll be extremely costly to generate one response.

Another thing that'd be really huge to test would be to regularly (e.g. once a year or quarterly) fine tune a per-user or per-organization model based on all our messages sent (so for fine tuning, prepare a dataset of Received Email > Reply), so it can really answer like we do, and reuse previous answers in a similar situation (much more flexibly than canned responses).

It won't be cheap (both the fine tuning and the fine-tuned model usage are quite pricey), and it'll will require some smart data cleaning to keep the training set as small but still comprehensive as possible, but this is the future of emails, so it'd be worth already experimenting with this. I've considered doing it on my end just to see what would come out, but Missive already has all the messages neatly structured for this, so it'd be the natural place to set this up.

If it works as well as I envision, it could make sense to even pay thousands of dollars per month for this (though hopefully the costs will come down quick), as the value of time saved will exceed that.

Erwin Mayer: The space is rapidly evolving and I anticipate that this will soon become possible. As Missive is not an AI company, we will likely integrate with products that offer this capability, similar to our current integration with OpenAI.

James McGlade: Yes, just waiting for the App store to approve the update, today, probably!

James Thurlow-Craig: Not really possible right now without spending significant resources. You can't retrain the whole GPT-3 corpus on a few canned responses or FAQ entry.